AI weather nowcasting · from satellite observations to interactive forecasts
Temporary demo pipeline for NVIDIA GTC Berlin 2026. It goes offline after the event. Read the blog post
Live · --:--
Forecast
--:--
Forecast initialised --:-- · Meteosat-12 FCI
Init
01 · AI Weather Nowcasting
From Satellite Observations to AI Forecasts
Nowcasting uses recent observations to predict weather over the next minutes to hours.
StormScope-EU uses six Meteosat-12 observations to predict twelve future satellite frames at 10-minute intervals, covering the next two hours.
Observed AI Forecast
−50 minInit +0+120 min
MeteosatNVIDIA AIEarthmoverViewer
Numerical weather prediction models take hours to run; AI nowcasting starts from what satellites see right now, fast enough to act on.
WHY IT MATTERS
Severe-storm warningsSolar power rampsAviationOutdoor events
02 · Earth Observations
Observing the Atmosphere from Space
Meteosat-12 observes Europe from geostationary orbit, approximately 36,000 km above Earth.
Its Flexible Combined Imager measures 16 spectral channels.
This demo shows two satellite composites, each formed by combining multiple spectral channels.
Two Satellite Composites
Observed +0 Frame
16spectral channels
10 minscan cadence
1–2 kmpixels, visible and infrared
03 · THE AI MODEL
StormScope-EU predicts the next two hours of clouds
A generative diffusion model from NVIDIA. It forecasts all 16 satellite channels directly in image space, learning how clouds move and grow instead of solving the physics.
INPUT · 6 OBSERVED FRAMES
−50−40−30−20−100
↓
StormScope-EU on GPU48 denoising steps · bf16
↓
+10 minThe new frame is fed back in. 12 rollouts reach +120 min.
Conditioned on orography and land–sea mask
Ensemble-of-experts denoising across noise levels
NVIDIA H200 · ~20 s per 10-minute step
04 · THE PIPELINE
Earth2Studio wires data, model and storage together
NVIDIA's open-source Python package for AI weather and climate. You can swap data sources, models and outputs without rewriting the pipeline.
1
FetchLatest 6 FCI frames from the EUMETSAT Data Store
2
PredictStormScope-EU rolls forward 12 × 10 min on the GPU
3
WriteUpload forecasts to Arraylake while GPU runs
4
ServeMap tiles served via Earthmover Compute
05 · DATA FLOW
From satellite to screen in ~20 minutes
Meteosat-12Scans Europe from orbit
10 min
EUMETSAT Data StoreProcesses and publishes the scan
+5 min
DownloadEarth2Studio fetches the latest frames
~2 min
Inference on NVIDIA H200 GPU saving to ArraylakePredicts 2 hours ahead
<4 min
Earthmover ComputeRenders map tiles straight from the data
~2 s
Your browserLoads all 13 frames and plays the loop
~12 s
EUMETSATNVIDIAEarthmover
06 · Earthmover Arraylake + Icechunk
Compute and Upload Together. Publish Once.
Forecast data is uploaded to Earthmover Arraylake while StormScope-EU generates subsequent predictions on GPUs.
Icechunk atomically publishes each complete forecast as a versioned snapshot.
Overlapped Uploads · Atomic Publication · Versioned History
07 · THE DATA LAKE
Every forecast run is a commit
Each 10-minute step uploads to Arraylake while the GPU computes the next one, and a single versioned Icechunk commit publishes all 12 at once. Icechunk atomically publishes each complete forecast as a versioned snapshot.
